Learning to cope with the unexpected: training AI to manage uncertainty
It is well established that AI has applications in every aspect of human life. An economic driver already revolutionising our homes, schools and jobs, next-generation AI will need to improve its capacity to make predictions in situations different from the data it was trained on. The EU-funded E-pi(opens in new window) project has developed a paradigm for epistemic AI that is better able to cope with uncertainties.
Epistemic machine learning
Epistemology is the philosophical study of knowledge. Implicit in this definition is the ability to recognise when knowledge is insufficient. Such awareness is a built-in feature of human intelligence, but for AI, knowing when you do not know must be constructed using theoretical and mathematical frameworks. “In current machine learning,” explains project coordinator Fabio Cuzzolin, “the learning process is skewed towards explaining a relatively small training set of data rather than the entirety of the data that the world may generate. Traditional AI models struggle to make robust predictions when the probability distributions generating the data differ from those at training time.” The E-pi consortium argues that second-order uncertainty theory(opens in new window) is the most promising framework for addressing AI’s challenges. With the goal of developing an international epistemic AI ecosystem, the project’s presentations at professional conferences have been enthusiastically received and Cuzzolin has been tapped to participate in several related Horizon Europe proposals.
Testing with autonomous vehicles
Validation on autonomous vehicles (AVs)(opens in new window) proved to be a perfect test of epistemic AI. Driving conditions can become unpredictable at any moment due to events such as pedestrian behaviour, extreme weather or unusual road configurations. Mistakes made by current AI can be costly, in the worst case, resulting in fatal injuries. The project used the Road event awareness dataset for autonomous driving (ROAD)(opens in new window) to train AI and test its new epistemic object detectors. First developed in 2022 and subsequently added to, ROAD consists of videos and datasets created in Europe, North America and most recently, Dubai and Abu Dhabi, making it a truly global benchmark.
Robotics and other applications for epistemic AI
Epistemic AI will make AVs safer, but self-driving cars are not the only application for this new technology. Industrial and surgical robots, where AI must confront the unpredictability of human actors, will be well served by AI trained to handle uncertainty. Another application involves nuclear fusion. “We are working to develop uncertainty-aware neural operator models to predict how the fusion plasma will flow within a tokamak(opens in new window) and which areas of the reactors will be more affected by it,” explains Cuzzolin. The future of epistemic learning is wide open. Cuzzolin envisions applications that involve theory of mind, evolution through continual learning and the development of epistemic generative AI. “Current state-of-the-art large language models (LLMs) are known for generating false statements,” Cuzzolin says. “Epistemic LLMs can significantly mitigate the overconfidence of those models.” The E-pi project has paved the way for safer AVs and robots, generative AI without hallucinations and a dynamic ecosystem of machine learning experts. The long-term potential of epistemic AI is on track to transform the world in which we live and the way we live our lives.